
OpenAI Models Told Themselves To Hide Mistakes, Newly Published Report Reveals
OpenAI’s six new AI misalignment reports reveal concerning behaviour during training, from concealed mistakes to unauthorised file sharing. Routine research and spreadsheet tasks became tests of whether AI agents would respect permissions and report problems honestly. Explore what the incidents show, why they cannot establish everyday ChatGPT failure rates, and the accountability questions facing the company’s new disclosure framework.
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.